One of the things I force myself to do is ask my AI to generate content, have me fully edit it, and then critique where it and I differ. Below is an example of one such interaction:
Bottom line
The generated draft reads like a carefully structured technical explainer. The published article reads like a personal, provocative newsletter essay.
The published version is more recognizably your voice: faster, more opinionated, more specific, more conversational, and less concerned with explaining every implication. It is also grammatically rougher in several places. The draft is cleaner and more logically complete, but it occasionally sounds over-structured and “written to a template.”
Structural snapshot
Image callouts excluded; draft headings counted as paragraph blocks.
| Measure | Generated draft | Published |
|---|---|---|
| Words | ~648 | ~614 |
| Paragraph blocks | 22 | 16 |
| Sentence-like units | 55 | 40 |
| Average words per sentence | 11.8 | 15.3 |
| Sentences of five words or fewer | 13 | 8 |
| Contractions | 0 | 2 |
| First-person references | 48 | 42 |
| Second-person references | 7 | 11 |
The published article is only about 5% shorter, but it has roughly 27% fewer sentences. In other words, it is more compressed, denser, and less staccato.
Major tone differences
1. The draft teaches; the published article declares
The draft carefully builds an argument:
- Route planning used to dominate.
- Detailed plans fail when reality changes.
- Destinations and fences are better inputs.
- Agents can discover the route.
- Human refinement still matters.
- Destinations must be verifiable.
- Fences prevent accidental operational problems.
That progression is textbook explainer structure. Each idea is introduced, defined, illustrated, and qualified.
The published article skips much of that scaffolding. It states the central position and expects the reader to keep up:
“I describe the output, I describe the boundaries, and I let my agent work out the middle.”
The definitions of “destination,” “boundaries,” and “fences” are removed. This makes the published version feel more confident and less instructional. It also leaves those terms slightly underexplained.
Effect: The draft sounds like a teacher making sure the argument survives scrutiny. The published version sounds like someone reporting a conclusion they have already reached.
2. Removing the headings makes the published version feel less generated
The draft had three conventional headings:
- “Today, I Set the Destination and the Fences”
- “What Actually Changed”
- “Conclusion”
They are clear, but generic. “What Actually Changed” and “Conclusion” especially resemble standard article-template scaffolding.
The published version removes all of them. Instead, it uses bold statements, images, and paragraph transitions to control the pace.
That is a strong tonal improvement for a newsletter. It produces a continuous personal essay rather than a formally segmented tutorial.
Effect: The draft feels organized from the outside. The published piece feels organized by the narrator’s train of thought.
3. The published title and subtitle are much more provocative
The published title is:
I No Longer Care How My Apps Get Built!
The subtitle says:
“I used to specify the input in obsessive detail. Now I specify the output and the fences, and let my agent find its own weird way there.”
This framing is considerably bolder than the body of the draft. “I no longer care” deliberately overstates the position; you clearly still care about boundaries, outcomes, taste, and refinement.
That overstatement is useful editorially. It creates tension the article can resolve.
The phrase “its own weird way there” also adds personality absent from the more neutral generated framing.
Effect: The publication presents a contrarian opinion. The draft presents a carefully qualified observation.
4. The published version is more autobiographically specific
The draft says:
“I even wrote about and recorded a video…”
The published version says:
“For the KIRUPA site, I even wrote about and recorded a video…”
That small addition matters. “The KIRUPA site” makes the example concrete and lived-in. The live link reinforces that this was an actual practice, not a hypothetical anecdote.
The PHP example is linked as well.
These proper nouns and references make the article sound more like a person drawing from their own history and less like a generalized essay about AI development.
5. The published version repairs a point-of-view wobble
The draft says:
“Usually a lot more than once when we learned more about user preferences…”
The article begins and continues as a first-person account, so this sudden “we” slightly blurs the narrator.
The published version changes it to:
“when I learned more about user preferences…”
That keeps ownership with the narrator and strengthens the memoir quality.
6. The agent becomes a character
The draft says:
“Then I stop planning and let it drive.”
The published version says:
“I stop planning and let my agent drive:”
“My agent” is clearer than “it,” but it also personifies the agent. The agent now:
- drives,
- picks paths,
- doubles back,
- throws work away,
- defines routes.
This strengthens the road-trip metaphor and makes the agent feel like a recurring actor rather than an abstract tool.
7. The published version trusts the metaphor more
The draft often translates the metaphor back into literal engineering language:
“course correct to meet users’ expectations”
The published version keeps the map language alive:
“course correct to meet the new destination”
Similarly, it expands:
“yet it arrives”
to:
“yet it arrives at the destination.”
This is slightly less idiomatic in places, but it preserves the article’s central visual vocabulary.
Effect: The draft prioritizes technical precision. The published version prioritizes thematic cohesion.
8. The human role changes substantially
In the draft, the human’s new work is primarily:
- defining the outcome,
- defining boundaries,
- verifying arrival,
- checking operational constraints,
- refining against something concrete.
The published version introduces a different center of gravity:
“This is where things like product sense and taste come into the picture in full force.”
That changes the thesis. The human is no longer mainly the constraint setter and verifier. The human becomes the source of judgment, aesthetic direction, and product intuition.
This is a meaningful improvement if the intended argument is about what remains uniquely valuable in human work. It also broadens the article beyond engineering into product design.
9. The draft is cautious; the published version is futurist
The draft contains an important operational caveat:
“Define a destination you cannot verify, and you are not doing output-focused development. You are just hoping.”
It also explains that fences prevent “it works” from silently meaning expensive, fragile, or local-only.
The published article removes that entire verification argument. In its place, it adds objections about:
- a billion concurrent users,
- enterprise readiness,
- compliance,
- bugs,
- data loss.
But instead of answering them with verification and constraints, it predicts:
“Longer term (like say, 3 months from now!) this would just be a detail we tell our LLM to handle for us.”
That is a major tonal shift.
| Draft | Published |
|---|---|
| Trust, but verify | Model capability will absorb more responsibility |
| Operationally cautious | Technologically optimistic |
| Timeless engineering principle | Deliberately time-sensitive prediction |
| Reassures skeptical engineers | Provokes skeptical engineers |
The published version is more likely to generate replies. The draft is more likely to survive as durable engineering advice.
10. Humor is concentrated differently
The draft’s humor is understated:
- “You are just hoping.”
- “until the bill shows up”
- the agent choosing PHP.
The published piece removes most of those understated jokes and puts its strongest humorous beat near the end:
“like say, 3 months from now!”
It also uses deliberately casual phrases such as:
- “some random bug”
- “its own weird way there”
- “a billion concurrent users”
The humor is broader and more topical. It supports the provocative newsletter tone rather than the teaching tone.
11. The published version addresses the reader more directly
The draft has approximately seven second-person references. The published version has around eleven.
The largest change is the new objection sequence:
“What if you need to handle a billion concurrent users? What about enterprise readiness and compliance?”
This turns the piece into a conversation with a skeptical reader. The draft anticipates objections through principles; the published version dramatizes the reader actually raising them.
Effect: The published version feels more interactive and debate-oriented.
12. The endings have different emotional shapes
The draft ends its argument with:
“The destination has not changed. The part I author has changed completely.”
That is reflective, balanced, and memorable. It closes the metaphor cleanly.
The published version removes it and places the scaling/LLM prediction after what had previously functioned as the conclusion.
As a result:
- The draft ends on personal transformation.
- The published version ends on technological acceleration.
- The draft closes the argument.
- The published version reopens the argument immediately before the call to reply.
That is less formally tidy, but more effective if engagement is the goal.
Sentence-level grammatical improvements in the published version
1. It removes a tautological aside
Draft:
“All of this planning and pre-work felt like a chore, and honestly a lot of it was.”
“Felt like a chore” followed by “a lot of it was” repeats essentially the same claim. The pronoun “it” is also slightly imprecise.
Published:
“All of this planning and pre-work felt like a chore.”
Cleaner and more confident.
2. It clarifies the driver
Draft:
“let it drive”
Published:
“let my agent drive”
The published noun phrase has a clear antecedent and reinforces the metaphor.
3. It simplifies an overly long refinement sentence
The draft explains refinement in a 37-word sentence:
“The destination gets sharper because I am reacting to something real instead of imagining it…”
The idea is good, but the sentence keeps appending explanations.
Published:
“So I look at it and I nudge until the destination gets sharper:”
This is shorter, more active, and visually sets up the next diagram.
The tradeoff is that it removes the valuable explanation of why reacting to something real is better.
4. It removes stacked modifiers
Draft:
“This all only works when…”
“This,” “all,” and “only” compete for emphasis.
The published version removes this sentence entirely, avoiding the construction.
5. It replaces excessive fragments with compound sentences
Draft:
“On the map, it looks inefficient. And it arrives.”
Published:
“On the map it looks inefficient, yet it arrives at the destination.”
Both are valid stylistic choices. The draft uses a dramatic fragment-like beat. The published version is grammatically more conventional and less theatrical.
Grammatical regressions in the published version
Several published sentences sound natural when spoken but are mechanically weaker than the draft.
1. “Details-oriented” should be “detail-oriented”
Published:
“less details-oriented”
The standard compound adjective is:
“less detail-oriented”
“Detail” modifies “oriented”; it normally remains singular.
2. “Where my attention was focused on” combines two constructions
Published:
“I just shifted where my attention was focused on”
This blends:
- “where my attention was focused”
- “what my attention was focused on”
Either is grammatical. Combining “where” with the final “on” is not.
The draft was cleaner:
“It moved where the specificity lives.”
An even simpler published formulation would have been:
“I just shifted my focus.”
3. “What the agent generates back” is not idiomatic
Published:
“What the agent generates back is usually about 70%…”
Natural alternatives are:
- “What the agent generates is usually…”
- “What the agent gives me back is usually…”
- “What comes back is usually…”
The draft’s “What comes back” is the cleanest of the three, although it does not explicitly name the agent.
4. “Make to will that into existence” creates a grammatical traffic jam
Published:
“the refinements you need to make to will that into existence”
This is technically decipherable, but “make to will” forces the reader to reparse the sentence. Two infinitive relationships collide:
- refinements you need to make
- in order to will something into existence
A cleaner structure would separate them:
“the refinements required to will it into existence”
The pronoun “that” is also less natural here than “it.”
5. “Your or your users’ expectations” lacks clean parallelism
Published:
“won’t meet your or your users’ expectations”
The intended meaning is clear, but the coordination is compressed awkwardly. “Your” first operates without an explicit noun and then appears again inside “your users’.”
More parallel:
“won’t meet either your expectations or your users’ expectations”
Or more concise:
“won’t meet your expectations or those of your users”
6. You normally reach a destination rather than meet it
Published:
“course correct to meet the new destination”
“Meet” works naturally with requirements or expectations. A traveler normally “reaches” a destination or “heads toward” one.
The sentence is preserving the metaphor, but the verb came from the draft’s earlier phrase “meet users’ expectations.”
More idiomatic:
“course-correct toward the new destination”
7. “Input-side” is unnecessarily hyphenated
Published:
“in the input-side of the problem definition”
Hyphenation is useful when the compound is attributive:
“input-side constraints”
Here, “side” is a noun followed by an “of” phrase:
“on the input side of the problem definition”
“On” also sounds more natural than “in.”
8. The “Longer term” sentence needs punctuation
Published:
“Longer term (like say, 3 months from now!) this would…”
The introductory phrase should be separated from the main clause. “Say” is also parenthetical and ordinarily takes commas.
Mechanically cleaner:
“In the longer term (say, three months from now!), this will…”
The published version’s punctuation contributes to its spontaneous voice, but the sentence currently reads as if it was transcribed from speech.
9. “Some random bug” and “some user” repeat the same vague determiner
Published:
“What if some random bug causes data loss for some user?”
This is conversational rather than grammatically wrong. Still, repeating “some” makes the sentence feel improvised.
Cleaner:
“What if a random bug causes data loss for a user?”
The original may be intentionally casual, especially beside the exaggerated billion-user example.
10. The introductory comma after “On the map” disappeared
Draft:
“On the map, it looks inefficient.”
Published:
“On the map it looks inefficient…”
The comma is not absolutely required for such a short introductory phrase, but including it improves readability and matches the natural pause.
Punctuation and formatting differences
Colons now carry much more structural weight
The published article repeatedly ends setup paragraphs with colons:
- the opening road-map premise,
- the completed plan,
- the failed plan,
- letting the agent drive,
- moving specificity,
- sharpening the destination.
This integrates the prose with the illustrations. The sentence before each image announces that the visual is the next part of the thought.
That is much better than treating diagrams as detached inserts.
Bold replaces headings
The published version bolds central claims such as:
- writing everything down first,
- describing output and boundaries,
- the scaling objection,
- the three-month prediction.
It also emphasizes product sense and taste.
This gives readers scanning anchors without interrupting the essay with generic headings.
Contractions make the voice less formal
The draft contains essentially no contractions. The published version uses forms such as:
- “wasn’t”
- “won’t”
That is a small change, but it contributes to the spoken, newsletter-like tone.
What the published version improves most
- It sounds more authored. Proper nouns, links, rhetorical questions, and slightly rough conversational syntax make it feel lived rather than constructed.
- It moves faster. Definitions and qualifications are cut unless they contribute directly to momentum.
- It has a stronger editorial posture. The title, product-sense argument, and three-month prediction give readers something to react to.
- It is better integrated with the diagrams. Colons and bold transitions make the visuals part of the prose.
- It trusts the audience. It does not explain every term or consequence.
What the generated draft does better
- Its core grammar is cleaner.
- It defines “destination” and “boundaries” concretely.
- It explains why iterative refinement works rather than merely stating that it happens.
- It includes the crucial requirement that arrival must be verifiable.
- It gives fences an operational purpose: cost, reliability, deployment, and durability.
- Its closing metaphor lands more cleanly.
- Its argument is more durable and less dependent on a near-term prediction about LLM capability.
The largest difference is therefore not stylistic but conceptual: the draft says verification and constraints make agent autonomy safe; the published article says product sense and taste make it useful, while rapidly improving models will absorb more of the remaining engineering responsibility. That is a genuine change in thesis, not merely an edit for length.
